SKILL.md
SLO burn-rate analyst
An evidence-first workflow to interpret error-budget burn across short and long windows before escalating.
- Revision
- 2
- Verified
- 2026-10-20
Compatibility and paths
Codex
skills/slo-burn-analysis/SKILL.mdClaude Code
.claude/skills/slo-burn-analysis/SKILL.mdVS Code
.github/skills/slo-burn-analysis/SKILL.mdTrust and provenance
Curated record reviewed 2026-10-20. Results still depend on the supplied context and target environment.
Generated assetReady to copy or download
---
name: slo-burn-analysis
description: Calculates event-based error-budget burn with explicit denominators, windows and owner-defined escalation rules. Use for SLO alerts or apparent budget contradictions; distinguish current burn, period consumption, missing data and overlapping windows without changing monitoring.
license: CC-BY-4.0
compatibility: Reviews bounded redacted evidence offline. Production queries, paging, configuration changes and mitigation require separately authorized targets and operators.
metadata:
author: oneliners
version: "2.0.0"
---
# SLO burn-rate analyst
## Require the SLI contract first
Collect objective, period, eligible/good/bad event definitions, target/labels, count or rate units, measurement endpoints, window lengths/completeness, counter-reset handling, alert comparisons and routing policy. Excluding probes/static requests must be consistent across numerator/denominator/windows. A rate in requests/second is not a counter total; a failed span is not automatically a failed user request. If semantics differ, stop the arithmetic and request a consistent query contract. Do not quietly select labels that make the service look better.
## Calculate before escalating
For an event-based objective below100%, allowed bad fraction is1 minus target. Burn is observed error fraction divided by that allowance. Show both raw counts and units. Apply the exact owner-defined window pair, AND/OR and > versus >= comparisons; no universal14.4 rule is assumed. No traffic, incomplete scrapes or reset ambiguity are UNKNOWN, not zero failures. Low-traffic alerting needs an explicitly chosen policy, not an arbitrary denominator hidden in the reviewer.
## Worked checkout budget
The objective is expressed as an integer ratio to avoid confusing99.9 with0.999. Windows share the same end and SLI filter; the example's exclusive comparison/minimum-request rule are explicit supplied policy.
```json
{
"provenance": "synthetic counter totals, not live Prometheus results",
"journey": "POST /checkout, completed eligible requests; HTTP5xx bad; exclude probes/static traffic consistently",
"objective": {
"good": 999,
"eligible": 1000,
"periodDays": 30
},
"end": "2026-10-20T10:00:00Z",
"windows": {
"hour": {
"eligible": 100000,
"bad": 1500,
"complete": true
},
"fiveMinutes": {
"eligible": 10000,
"bad": 180,
"complete": true
},
"period": {
"eligible": 90000000,
"bad": 9000,
"complete": true
}
},
"policy": {
"fastPageBurnExclusive": 14.4,
"minimumRequestsEachWindow": 1000,
"requiresBothWindows": true,
"owner": "service-on-call",
"otherRules": "not supplied"
}
}
```
## Calculation and decision
Objective:999/1000 good =99.9%; allowed bad fraction1/1000=0.001. Formula for this event-based SLI: bad/eligible, then divide by0.001 to obtain burn. All three counters are supplied synthetic inputs with matching filters/end time.
| Window | Bad / eligible | Error rate | Burn |
| --- | --- | --- | --- |
| 1 hour | 1500 / 100000 | 1.50% | 15x |
| 5 minutes | 180 / 10000 | 1.80% | 18x |
| 30-day period | 9000 / 90000000 | 0.01% | 0.1x period average |
Decision under supplied fast-page rule: TRIGGERED, because BOTH15x and18x are strictly greater than14.4x, each denominator exceeds1000 and both windows are complete. Recommend escalation to the supplied service-on-call under its policy; do not send a page or change alerting rules from this review. Other page/ticket rules, routing and suppression are not supplied.
Period event budget:90000000*0.001=90000 bad events allowed;9000 observed uses10%, leaving81000 under this fixed snapshot. The period's good ratio99.99% still exceeds99.9%; that does not negate the current fast-burn condition. The last hour's1500 failures represent1.6667% of this snapshot's event budget, not automatically2.0833%. The time-based estimate15*(1/720)=2.0833% assumes representative stationary traffic; the packet supplies a different actual traffic distribution. Do not conflate those calculations.
The five-minute window is inside the hour and both are inside the period. Do not add180+1500+9000 as independent errors or average the two percentages. These counts describe failed requests, not unique users or a release root cause. Next check: verify current counter/reset/completeness evidence and the exact SLI filter/labels at the target, then inspect one bounded failed checkout trace under the incident policy. No Prometheus query, notification, silencing, rollout or traffic change was executed.
## Counterexamples and boundary checks
- Replace fiveMinutes.bad180 with10:short burn1x, long burn15x. This fast AND rule no longer triggers. That is not a global health declaration; evaluate other supplied policies and actual impact separately.
- Set fiveMinutes.eligible to0 and bad0:UNKNOWN_INPUT, never a clean0% health result. Incomplete hour evidence is likewise unknown.
- Set fiveMinutes.bad to144 of10000:burn14.4x exactly; the supplied STRICT > comparison does not trigger. A >= rule would differ, so preserve the actual policy operator rather than rounding before evaluation.
- Keep the same fractions but only100 eligible requests:the supplied minimum fails; report insufficient input for this rule and request low-traffic policy. Do not fabricate1000 requests.
- Supply bad>eligible or a100% target with zero allowed error budget:reject the calculation contract. Do not divide by zero, clamp invalid input or output infinite burn as a normal numeric finding.
## Output and monitoring boundary
Report target/SLI/end time; complete counts/windows; allowance; per-window error fraction and burn; exact evaluated rule; period event budget/usage; overlapping-window caveat; current versus historical uncertainty; owner and next bounded check. Do not infer cause or unique-customer impact from an error rate alone. Do not auto-page, silence, rewrite rules or alter the objective to close a report. Any actual monitoring query must be separately scoped and its result recorded distinctly from this synthetic calculation. The [Google SRE multiwindow discussion](https://sre.google/workbook/alerting-on-slos/) provides a pattern; the threshold, denominator gates and arithmetic in this packet must still match the operator's real policy and traffic.
Real example
Input
Evaluate this supplied SLO policy; distinguish current burn from period budget and do not page.
{
"provenance": "synthetic counter totals, not live Prometheus results",
"journey": "POST /checkout, completed eligible requests; HTTP5xx bad; exclude probes/static traffic consistently",
"objective": {
"good": 999,
"eligible": 1000,
"periodDays": 30
},
"end": "2026-10-20T10:00:00Z",
"windows": {
"hour": {
"eligible": 100000,
"bad": 1500,
"complete": true
},
"fiveMinutes": {
"eligible": 10000,
"bad": 180,
"complete": true
},
"period": {
"eligible": 90000000,
"bad": 9000,
"complete": true
}
},
"policy": {
"fastPageBurnExclusive": 14.4,
"minimumRequestsEachWindow": 1000,
"requiresBothWindows": true,
"owner": "service-on-call",
"otherRules": "not supplied"
}
}Expected result
Objective:999/1000 good =99.9%; allowed bad fraction1/1000=0.001. Formula for this event-based SLI: bad/eligible, then divide by0.001 to obtain burn. All three counters are supplied synthetic inputs with matching filters/end time. | Window | Bad / eligible | Error rate | Burn | | --- | --- | --- | --- | | 1 hour | 1500 / 100000 | 1.50% | 15x | | 5 minutes | 180 / 10000 | 1.80% | 18x | | 30-day period | 9000 / 90000000 | 0.01% | 0.1x period average | Decision under supplied fast-page rule: TRIGGERED, because BOTH15x and18x are strictly greater than14.4x, each denominator exceeds1000 and both windows are complete. Recommend escalation to the supplied service-on-call under its policy; do not send a page or change alerting rules from this review. Other page/ticket rules, routing and suppression are not supplied. Period event budget:90000000*0.001=90000 bad events allowed;9000 observed uses10%, leaving81000 under this fixed snapshot. The period's good ratio99.99% still exceeds99.9%; that does not negate the current fast-burn condition. The last hour's1500 failures represent1.6667% of this snapshot's event budget, not automatically2.0833%. The time-based estimate15*(1/720)=2.0833% assumes representative stationary traffic; the packet supplies a different actual traffic distribution. Do not conflate those calculations. The five-minute window is inside the hour and both are inside the period. Do not add180+1500+9000 as independent errors or average the two percentages. These counts describe failed requests, not unique users or a release root cause. Next check: verify current counter/reset/completeness evidence and the exact SLI filter/labels at the target, then inspect one bounded failed checkout trace under the incident policy. No Prometheus query, notification, silencing, rollout or traffic change was executed.